Given a sequence of DNA nucleotide bases, the task of gene prediction is to find subsequences of bases that encode proteins. Reasonable performance on this task has been achieved using generatively trained sequence models, such as hidden Markov models. We propose instead the use of a discriminitively trained sequence model, the conditional random field (CRF). CRFs can naturally incorporate arbitrary, non-independent features of the input without making conditional independence assumptions among the features. This can be particularly important for gene finding, where including evidence from protein databases, EST data, or tiling arrays may improve accuracy. We evaluate our model on human genomic data, and show that CRFs perform better than HMM-based models at incorporating homology evidence from protein databases, achieving a 10% reduction in base-level errors.